Certainly it is an interesting phenomenon.
It may be caused by dendritic growth within the moulding compound surrounding the leadouts of the Blue LEDs.
If the two leadout wires are close together, and there is some moisture content, or uncured resin or activator within the moulding, then the potential gradient may be sufficient, over a long period with applied voltage bias, to cause dendrites to form within the moulding by an electrochemical migration process.
Eventually, these dendrites may result in a resistive conductive channel across the LED, which will divert the current that was intended for the LED. When an increased current is supplied, this blows away part of the dendrite (like a fuse blowing), allowing the LED to work again. However, when the current is reduced, the dendrites grow back again so that the LED becomes intermittent and then stops working.
Another mechanism is the growth of tin whiskers. However, these are much smaller diameter, and will probably cause a direct short, which would have been detected by your multimeter tests. Also I believe that the tin whiskers would grow regardless of the applied voltage bias.
The dendrites may be visible (assuming a transparent moulding) with your microscope, as a dark channel between the leadout frame.
https://nepp.nasa.gov/whisker/background/index.htmHope this helps